Council Post: The Next AI Breakthrough Isn’t Smarter Models—It’s Something The Industry Forgot To Build

Juan Graña, CEO and founder, Neurologyca, a leader in Human Context AI.

getty

A few years ago, most interactions with AI were impressive but shallow. The “intelligence” part of AI was more of a blunt stick than a sharp blade, but the automation and split-second responses were so novel that it caught on quickly.

Its popularity, for businesses and individuals, meant that conversations around AI quickly became about scale and capacity—bigger models, more training data and faster infrastructure to support human-machine interaction. Systems can now generate content, write code and carry out increasingly complex tasks with minimal input, or no user input at all in the case of agentic AI.

Each new iteration of AI has pushed the boundaries further, requiring more compute, data center capacity and power. But in focusing so heavily on how powerful AI can become, we’ve overlooked a more fundamental question—how well does it work for the people actually using it?

Spend time with any AI model or AI-powered system, either at work or at home, and a pattern emerges. You begin with a clear objective, but as you move through the interaction—which might span countless prompts and replies—your intent begins to shift. You might refine what you’re asking, hesitate over a decision or change your direction entirely based on new information, all while the system continues to move forward based on what it last understood. The system is busy interpreting rather than understanding, and that’s the missing layer: human context.

Why Scaling AI Solves The Wrong Problem

As AI systems have taken on more responsibility, they’ve also moved further away from the original point of instruction. What starts as a clear prompt can quickly turn into a chain of decisions, actions and outputs that unfold over time.

In theory, this is where AI becomes most valuable. It can handle complexity, reduce manual effort and carry intent forward without constant intervention. That’s the promise, but in practice, it’s where the biggest problems start to emerge. The further a system gets from that initial input, the more it has to rely on its own interpretation of what the user meant, rather than what they mean in the moment.

You can see this in everyday use. A hiring manager refining a shortlist might shift their criteria halfway through reviewing candidates. A developer iterating on a piece of code might realize the original approach no longer fits the problem. A student working through a lesson might lose focus and need to adjust their pace.

In each case, the user’s needs and intent are evolving in real time, shaped by the context of their lived experience. Most AI systems aren’t designed to track those changes, no matter how powerful or impressive the model. They respond to what was said rather than how the person is responding. Over time, that creates a sort of “AI drift,” where the system is moving forward while no longer being fully aligned with the person it’s meant to be serving.

The Need For Human Context

Addressing this AI drift means developing a way for AI systems to understand how people are actually responding as an interaction unfolds, rather than just referring back to the last prompt or guessing based on re-prompts.

In most cases, the signals are already there. You can often see when someone slows down, revisits a step, abandons a line of thinking or leans into a particular direction. These are the natural rhythms of how people think, decide and adapt in real time. Systems should capture those signals and carry them forward, making more helpful suggestions instead of treating each interaction as a fresh input iterating on the last. This is the human context layer that I believe is needed.

One approach is to treat those behavioral signals as a form of structured data—anonymized—that can be read and used by the system itself. Instead of relying solely on prompts, the system could build a live picture of how the user is engaging: where they’re confident, where they’re hesitating and how their focus is shifting. That context should persist across the interaction, even setting a baseline for future interactions.

There’s also an opportunity (with the user’s consent) to share that anonymized data so that models can improve their interactions with other people in similar situations, environments or lines of work—whether it’s self-improvement, AI-assisted learning in high school or aiding decision-making at the office. This new layer of human context can help transform AI into a true supporting co-pilot rather than a bot executing instructions.

In practice, organizations don’t need to rebuild their AI systems from scratch to add this capability. A human context layer can sit between the user interface and the AI model or agent, continuously translating consented behavioral signals—such as interaction patterns, hesitation, attention shifts or changes in engagement—into structured, machine-readable context.

Rather than telling the system that someone is definitively “confused” or “stressed,” these signals can be expressed as dynamic indicators with confidence levels and passed to the AI alongside the task. The application can then define how the system should adapt: slowing the pace, simplifying an explanation, asking for clarification, surfacing alternatives or escalating a decision to a human. This turns human context from an observation into an input that can actively shape the AI’s next action.

​Closing Thoughts

When AI systems begin to respond to human behavior as it unfolds, the interaction can change from something the user needs to manage to something that meets them where they are. The small signals that usually go unnoticed—like lingering on a step, second-guessing a choice or circling back to an earlier idea—start to shape how the system responds in real time. The system doesn’t need to be constantly re-prompted, and the platform can adjust its pace, level of detail and framing based on how the user is actually moving through the task.

Over time, this can reduce the need to “steer” the system, because it is already tracking the direction you’re heading in. It’s in the passenger seat, not tailing behind in another vehicle trying to keep up. ​​​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?